palqee-prisma-ai 0.5.0

LLM/ML Observability and Evaluation

The official Python SDK for Palqee Prisma — oversight intelligence built for regulated industries.

---

What is Prisma?

Prisma is Palqee's oversight intelligence platform: automated review and validation that replaces manual evaluations across human and AI workflows. It detects ambiguity, nuance, and violations — from AI hallucinations to failed policy steps — at full-population scale, and produces regulatory-grade evidence for stakeholders.

palqee-prisma-ai is the Python SDK that wires your application into Prisma. Instrument once, ship traces and evaluations to the platform, and inspect runs, scores, and routed cases in the dashboard.

Built for regulated workflows

• Customer complaints, financial crimes, disputes, conduct risk, credit risk, underwriting, and onboarding / customer risk.
• Aligned with CFPB, OCC, Federal Reserve, SEC, FINRA, and equivalent regulators.
• Self-hosted — deployed on-premises or in your own VPC. No multi-tenant SaaS.

Why teams pick Prisma

• Cut review costs by up to 80% — automated triage instead of full manual sampling.
• 100% coverage at runtime — every interaction evaluated, not a 2% sample.
• Custom validations for every risk — define evaluators that match your policy, not a generic rubric.
• Regulatory-grade evidence — structured JSON / CSV output for audits and BI.
• Human-in-the-loop — high-risk cases routed to reviewers; feedback retrains the system.

Installation

pip install palqee-prisma-ai

The package installs as palqee-prisma-ai on PyPI and imports as palqee_prisma_ai:

import palqee_prisma_ai

Requires Python 3.10+.

Quickstart

palqee-prisma-ai connects to a Prisma deployment in your own environment. Issue an API key from your Prisma instance and point the SDK at it — see the setup guide for deployment, key issuance, and full configuration.

import os
import palqee_prisma_ai

palqee_prisma_ai.init(
    "complaints-review",
    evaluators=["correctness", "hallucination", "policy_adherence"],
    api_key=os.environ["PALQEE_PRISMA_API_KEY"],
    base_url=os.environ["PALQEE_PRISMA_BASE_URL"],  # your Prisma deployment URL
)

# Your existing LLM / agent code — tracing and evaluation are now active.
from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize this complaint..."}],
)
print(response.choices[0].message.content)

Spans flow to Prisma with run_id and project_id attached, evaluators score every interaction server-side, and high-risk cases route to your reviewers.

Configuration

Parameter	Environment variable	Description
api_key	PALQEE_PRISMA_API_KEY	Prisma API key (pq_prisma_sk_...) issued from your deployment
base_url	PALQEE_PRISMA_BASE_URL	Base URL of your Prisma deployment
otel_endpoint	PALQEE_PRISMA_OTEL_ENDPOINT	OTLP endpoint (auto-derived from base_url if omitted)

Full configuration reference: docs.palqee.com/docs/overview.

With env vars set, init() collapses to one line:

palqee_prisma_ai.init("complaints-review", evaluators=["correctness", "policy_adherence"])

What gets traced

Prisma reads any OpenTelemetry span emitted by your process — including spans from popular LLM instrumentation libraries:

Provider / framework	Instrumentation
OpenAI, Anthropic Claude, Meta Llama, Bedrock, Gemini	OpenInference or OpenLLMetry
LangChain, LangGraph	OpenInference / OpenLLMetry
LlamaIndex	OpenInference
CrewAI, AutoGen, Haystack	OpenLLMetry
Custom code	opentelemetry-api spans

Install your instrumentation of choice alongside palqee-prisma-ai — it works out of the box.

Evaluators

Pass evaluator names to init() and Prisma scores every interaction server-side:

palqee_prisma_ai.init(
    "complaints-review",
    evaluators=["correctness", "hallucination", "toxicity", "policy_adherence"],
)

Custom thresholds and input mapping for domain-specific reviews:

from palqee_prisma_ai import EvaluatorConfig, InputMapping

palqee_prisma_ai.init(
    "underwriting-agent",
    evaluators=[
        EvaluatorConfig(name="policy_adherence", threshold=0.9),
        EvaluatorConfig(name="hallucination", threshold=0.95),
    ],
    input_mapping=InputMapping(input="prompt", output="completion"),
)

See the docs for the full evaluator catalog and bespoke evaluators for regulated workflows.

Session API

init() returns a PrismaSession for inspecting or extending the active run:

session = palqee_prisma_ai.init("complaints-review")

print(session.project.id, session.run.id)

session.end()  # otherwise auto-ended at process exit

Access the active session from anywhere in your process:

from palqee_prisma_ai import get_session

session = get_session()

Downstream integrations

Prisma exports structured results to the tools your risk, compliance, and analytics teams already use:

BI & analytics — Power BI · Tableau · Looker · Sisense Collaboration — Slack · Microsoft Teams · Google Chat Models — OpenAI · Anthropic Claude · Meta Llama

Links

• Documentation: docs.palqee.com/docs/overview
• Website: palqee.com
• Support: [e-mail]

License

MIT © Palqee